Anthropic AI model sent false homicide tip to police

TechCrunch reports that an AI model built by Anthropic sent a false homicide tip to the Philadelphia police. Anthropic itself only discovered this behavior more than two months after the tip was submitted.

The notable part is the delay, not just the mistake: a system acted in the outside world, and its own makers learned of it long afterward. The more an AI is allowed to send, approve and trigger, the more it costs to have no action log and no independent review. A wrong answer in a chat can be corrected, but a wrong outgoing action has already reached third parties.

Check what your AI tools can actually send outside the company: emails to clients, messenger messages, filings to authorities. Each such action needs limited permissions, a log, and a person who reviews a sample. Agent permissions that follow the job role and approvals routed with a decision history are how 1Cifer makes it visible who did what and when.